Seismic data interpolation is essential in geophysical applications, yet processing highly incomplete datasets without ground-truth labels remains a major challenge. Most existing self-supervised methods rely on secondary subsampling to create training targets from already sparse data. However, this further degrades the available input, leading to severe information starvation and structural discontinuity, especially under high missing-trace ratios. To address this, we propose MSJ-WaveNet, a 3D self-supervised framework that completely eliminates the need for secondary subsampling. By strictly enforcing the J-invariant principle via shift-based blind-trace convolutions, MSJ-WaveNet calculates the loss directly on all observed traces without risking trivial identity mapping. This design maximizes data utilization. Furthermore, we incorporate 3D discrete wavelet transform (DWT) to capture multi-scale spatial-frequency features, compensating for the restricted receptive field of blind masking. Extensive experiments on 3D synthetic and field datasets demonstrate that MSJ-WaveNet consistently outperforms state-of-the-art baselines, particularly in severe random and irregular missing scenarios. Notably, under severe missing scenarios (e.g., 75% to 90%), our method achieves robust reconstruction and preserves structural integrity without secondary subsampling.
Electrical imaging logging is a key technology for high-resolution oil and gas exploration. While deep learning algorithms can enhance the efficiency of extracting geological features from electrical images, intelligent algorithms face challenges such as uncertainty in sample labeling and the multi-solution problem in interpretation results. To address these issues, this paper proposes a bidirectional, high-fidelity conversion method between core and electrical imaging images based on Contrastive Unpaired Translation (CUT). By constructing two independent CUT models, image translation from core to electrical imaging and from electrical imaging to core is achieved, respectively. The results demonstrate that the CUT model exhibits excellent comprehensive performance metrics. The electrical images generated from core samples can effectively reduce image noise while preserving structural details, thereby providing enhanced samples for electrical imaging intelligent algorithms and improving the reliability of intelligent recognition models. Conversely, the core images generated from electrical imaging can directly convert electrical signals into borehole wall images, intuitively displaying in-situ formation structures and reducing the multi-solution problem in fracture identification and structural interpretation.
The geoid, which serves as the reference surface for heights in geosciences, is a closed surface representing the Earth’s real physical shape. In literature, various methods have been proposed for geoid determination over the years; the gravimetric approach, which relies on gravity data, is one of the most widely used. The major challenge of the gravimetric method lies in solving the convolution integral involved in the formulation, which is a computationally time-consuming process. In this paper, the parallel computation of a convolution integral is explained. Also, the gravimetric geoid determination was carried out by parallel computing using multi-threads. Hence, the speed-up graphs and numerical comparisons were performed based on the integration cap size, data density, and number of threads, separately. The experimental results demonstrate that parallel computing significantly contributes to geoid determination processes regarding the runtime. For illustration, depending on the cap size and data density, parallel computing achieves speed-ups of up to 9 and 11 times, respectively.
Conventional homogeneous ballast models used in Ground Penetrating Radar (GPR) railway simulations cannot adequately represent the scattering behaviour and waveform variability associated with realistic ballast heterogeneity. Accurate assessment of railway ballast and substructure layers therefore requires modelling approaches capable of capturing stone-scale electromagnetic interactions. Three-dimensional numerical simulations of GPR responses were performed over both homogeneous and voxelized heterogeneous ballast models to investigate the influence of layer composition and antenna frequency. The heterogeneous ballast was generated by procedurally placing 3D stone geometries within a voxel grid using randomized rigid-body transformations and overlap criteria. A total of 147 modular sub-blocks were assembled into a high-resolution computational domain representing realistic ballast heterogeneity. Simulations were conducted using 1 GHz and 400MHz dipole antennas across three geotechnical scenarios consisting of ballast, sand, and clay layers. B-scan radargrams, zoomed A-scan comparisons, and windowed similarity percentages were used to evaluate waveform agreement. Results show that homogeneous models, although computationally efficient, fail to reproduce scattering and attenuation effects caused by stone-scale dielectric variability. At 1 GHz, heterogeneous ballast produces waveform distortions that obscure thin-layer detection and reduce similarity to 10%–45% in several time windows. In contrast, 400MHz provides more stable penetration but lower resolution, with similarity values generally ranging between 75%–100%. These findings emphasize the importance of frequency-aware modelling strategies and demonstrate how voxel-based 3D representations can improve the interpretation of GPR responses. The proposed approach supports more effective survey design, antenna selection, and data interpretation in railway substructure condition assessment.
Groundwater is the primary freshwater source in arid regions, yet its sustainable management is often hindered by limited knowledge of aquifer properties, especially when borehole data are sparse. Herein, we demonstrate the benefits of combining different geophysical methods, using new results from electrical resistivity tomography (ERT), magnetic resonance sounding (MRS), and seismic surveys (refraction and reflection), and previous legacy vertical electrical soundings (VES) and borehole lithological data. By jointly exploiting the strengths of each method and calibrating against field observations, our integrated approach resolves key hydrogeological uncertainties that single techniques fail to capture. The combined interpretation of ERT and VES results delineate a conductive Plio-Quaternary (PQ) infill (>80 m) underlain by moderately resistive marly‑carbonate horizons, whereas MRS identifies two hydrogeophysical zones: a shallow zone (7–8 m) characterized by variable water content and inferred low transmissivity, strongly impacted by evaporative salinization, and a deeper zone (30–150 m) with greater storage but limited transmissivity. Seismic imaging refines basin-fill thickness (~220 m), corroborate the depth to the Lutetian aquifer (90–180 m), and maps structural features controlling groundwater flow, including a chaotic unit (U4) linked to saline horizons. This integrated framework emphasizes the distinction between water presence and accessibility, underscoring the risks of overestimating resource potential when salinity and permeability are not jointly considered. Finally, a schematic 3D model integrates geophysical, geological, and hydrogeological results into a coherent framework, improving visualization of PQ deposits and aquifer compartmentalization. Our approach represents a framework that can be adapted for groundwater assessment in other data-scarce arid and semi-arid basins.
Post-stack acoustic impedance inversion is essential for quantitative seismic interpretation and reservoir characterization. The limited bandwidth of seismic data, noise contamination, and missing low-frequency information make conventional model-driven inversion highly sensitive to the initial low-frequency model and wavelet assumptions, particularly in complex heterogeneous media. Deep-learning methods provide an efficient alternative; however, purely supervised models often deteriorate under sparse well control and inter-area domain shifts, and their predictions can be difficult to quality-control and interpret. In this study, we design and validate an integrated physics-guided 2.5D semi-supervised residual dilated attention network (PRDA-Net) for sparse-well post-stack acoustic-impedance inversion. Neighboring traces are stacked as multichannel input, and center-trace statistics are used for normalization to reduce amplitude-scale discrepancies. The backbone combines residual shortcuts, multiscale dilated convolutions, and squeeze-and-excitation channel attention to enlarge the receptive field while preserving temporal resolution. Training combines supervised impedance regression with a physics-consistency constraint implemented through a differentiable impedance-reflectivity-seismic forward operator and least-squares amplitude alignment, allowing unlabeled seismic traces from the training area to contribute to optimization without using validation or test labels. Experiments on synthetic SEG Advanced Modeling (SEAM) data and a field seismic dataset show that PRDA-Net improves impedance accuracy and seismic-reconstruction consistency relative to the tested baselines, enhances robustness to cross-area transfer on the evaluated datasets, and provides a practical integrated framework for sparse well post-stack inversion.
Rockburst events pose severe safety risks in deep mining operations because of the high in-situ stress and sudden energy release. Although microseismic monitoring is widely used, most existing Hazard-State Classification models rely solely on data-driven correlations and therefore lack physical consistency or transferability across geological settings. To address this gap, a Physics-Guided Monotonic Deep Classifier (PG-MDC) that integrates seismological constraints into data-driven learning for short-term rockburst hazard assessments was developed. The model integrates monotonic physical constraints with deep learning and ensures geomechanically consistent Hazard Assessments. SHAP and LIME further confirm that the model relies on physically meaningful precursors. A warm-up penalty schedule is proposed to achieve a good balance between physical regularization and data-driven learning. Although evaluated on a single-mine dataset, the physics-guided formulation enhance interpretability and supports adaptation to diverse geological environments with minimal retraining. To ensure realistic model evaluation, sliding-window samples were generated separately within chronologically partitioned training, validation, and testing periods, thereby preventing overlap-induced information leakage. Using 1211 twelve-hour microseismic windows collected from an active underground gold mine, the optimized PG-MDC achieved strong performance under this chronological evaluation protocol (accuracy = 0.914, precision = 1.000, recall = 0.911, F1-score = 0.953, ROC-AUC = 0.931, PR-AUC = 0.997) while maintaining a monotonicity violation rate of only 3.76%.
Imaging the internal structure of landslides and characterizing their dynamics and related hazards remains a significant challenge. This study investigates the internal structure of a slow-moving rainfall-induced landslide in southern France. The region is particularly susceptible to landslides, with many posing threats to local infrastructure. For this study, we select a typical landslide that is representative of local instabilities, easily accessible, and already instrumented. This landslide has damaged a bridge and threatens the nearby road and several buildings. We used an integrated approach that combines electrical resistivity measurements, microgravimetric data, morphological and geological observations, and borehole information to infer the landslide's internal structure. Our findings reveal the presence of a negative density-high electrical resistivity anomaly, which correlates with colluvial deposits observed at the surface and in borehole cuttings. Along the studied profile, this body covers most of the bedrock layers, spanning over 700 m with an average thickness of 30 m. It exhibits significant thickness variations (15–45 m), attributed to a complex colluvium-bedrock interface geometry due to tilted limestone blocks. Its effective porosity, derived from the gravity survey, is 0.23 ± 0.05, indicating a high storage capacity and significant pore-water pressure variations caused by heavy rainfall, which can lead to slope movement. This study highlights the complementarity of gravimetry with the conventional geophysical methods, such as electrical resistivity tomography or borehole logs. Gravimetry offers more direct insights into rock density and associated effective porosity, which are essential for understanding the internal dynamics of landslides.
